MolDeTr — chemistry-informed deep learning for ¹H NMR multiplet detection

MolDeTr is a 1D Deformable-DETR that reads a ¹H NMR spectrum window and returns the spin systems in it directly: for each group of equivalent protons it gives the chemical shift (δ), the coupling (J), the proton count, and the line width — in one forward pass, with no prior structure and no iterative fitting.

• Paper: Analytical Chemistry, 2026
• Code: https://github.com/smidooo/MolDeTr
• Canonical release (weights + data): Zenodo 10.5281/zenodo.21217101. This repo mirrors the checkpoint from that deposit for convenience — Zenodo is authoritative.

What's here

One file: model_spin_system_ABCDEFG_exp2.pth (~974 MB), the trained checkpoint. It is byte-identical to the file in the Zenodo deposit (MD5 faf842d1a1d8beae67e0544e28f226b5).

Usage

The model is custom (a 1D detection transformer), so it runs through the repo code rather than a standard transformers pipeline:

git clone https://github.com/smidooo/MolDeTr && cd MolDeTr
pip install -e .
huggingface-cli download smidooo/moldetr model_spin_system_ABCDEFG_exp2.pth --local-dir moldetr/model
python scripts/predict.py --demo          # or: python app.py  (Gradio Detect + Simulate app)

See the repository README for the input contract (a 6144-point, 5.12 points/Hz, 1200 Hz window) and the interactive app.

Benchmark

On the experimental test set (13 ROIs across 12 spectra, 80–600 MHz, vs. ground truth): median |Δδ| 0.89 Hz, median |ΔJ| 0.20 Hz, and 93.5 % proton-count accuracy.

License & citation

Apache-2.0. If MolDeTr helps your work, please cite the paper.